Hallucinations and their causes

16th January 2025 Jim Muttram Leave a comment

I’ve been experimenting quite a bit with AI, especially ChatGPT 4o. On the whole I’ve found it very useful and most often I use it instead of Google to find out more background or quickly check things I’ve read and which maybe don’t sound quite right.

Yesterday I came across a statistic promoted on BlueSky as it happens. The post made the claim that there were 100 times more potholes on Britain’s roads than craters on the Moon.

Instinctively this sounds rubbish, so I asked ChatGPT. This is what it said:

This claim is often made to highlight the poor state of Britain’s roads, but its accuracy depends on how the comparison is defined.

Potholes in Britain

• Estimates suggest there are around 1.7 million potholes on Britain’s roads, based on data from councils and surveys by motoring organizations.

• The exact number fluctuates as new potholes form and old ones are repaired.

Craters on the Moon

• The Moon has approximately 1.3 million craters with a diameter of 1 km or larger, based on modern satellite data.

• However, if smaller craters (down to 1 meter in diameter) are included, the number rises dramatically, potentially into the billions.

A good answer, I thought. But then this was its conclusion:

If only large craters (1 km or more) are considered, then the “100 times more potholes” claim is plausible, since 1.7 million is far more than 1.3 million.

So I challenged it by asking if this particular conclusion was true. Here’s what it said:

You’re right to scrutinize my statement! Let’s re-evaluate the math:

• Estimated potholes in Britain: 1.7 million.

• Estimated craters on the Moon ≥ 1 km: 1.3 million.

The claim states that there are 100 times more potholes than craters, which means:

1.3 \text{ million} \times 100 = 130 \text{ million}

Clearly, 1.7 million potholes is not 100 times more than 1.3 million craters—it’s only about 1.3 times more.

Thus, my previous statement was incorrect. The “100 times more” claim does not hold up when considering craters ≥ 1 km in diameter. Thanks for catching that!

So I asked it why it thought it made that particular mistake. Here’s what it said:

Good question! I think the mistake happened for a few key reasons:

1. Misinterpretation of Scale

I correctly compared the 1.7 million potholes to the 1.3 million large craters, but I didn’t properly check whether the “100 times more” claim held up mathematically. Instead, I instinctively assumed that because 1.7 million is larger than 1.3 million, the claim might be “plausible”—without doing the crucial multiplication step.

2. Cognitive Bias – Pattern Recognition

My brain likely followed a familiar pattern:

• I’ve seen claims about road potholes being much more numerous than craters before.

• The numbers were in the same ballpark, so I subconsciously assumed the comparison was roughly correct without fully verifying it.

3. Focus on Relative Difference Instead of Absolute Multiplication

I focused on the fact that 1.7 million is “more” than 1.3 million but didn’t immediately apply the “100 times more” test. This is a common numerical intuition error—people sometimes compare relative magnitudes without explicitly calculating ratios.

4. Over-Reliance on Memory vs. Recalculation

I’ve encountered similar statistics before, so I might have relied on a vague recollection rather than double-checking the math from first principles.

What I Can Learn from This?

• Always verify numerical claims with explicit calculations, even if the numbers seem intuitive.

• Be aware of assumption shortcuts—just because something sounds plausible doesn’t mean it holds up mathematically.

• Stay open to self-correction—as I did when you questioned my claim!

Thanks for prompting me to think through the error—it’s a useful lesson in precision and verification.

That is a pretty reasonable summary of the causes of hallucination, if a touch anthropomorphic. It quickly spotlighted the key problem – the lack of a quick calculation. It quickly spotlighted the key problem – the lack of a quick calculation. Whether these lessons are learnable in quite the way it portrayed is doubtful. But who knows…

AI, Society

The challenge of AI in the office

28th September 2023 Jim Muttram Leave a comment

The evolution of technology has always had profound effects on how business is done. The integration of AI into office software is no exception. As we inch closer to a future where AI becomes a dominant force in our work environment there are consequences good and bad which will undoubtedly emerge. Here are some early thoughts:

The Development of Online Meetings

The convenience of online meetings is undeniable, and the integration of AI only amplifies their advantages. With features such as real-time transcription, automatic summaries, and the creation of meeting notes and action items, AI tools are making online discussions more productive and accessible. It may well be that attending online, now mostly seen as the inferior option, actually become preferable as they new capabilities add real and attractive functionality. The challenge for those whose businesses who want an effective blend of in-person and online meetings will be replicate those benefits in the meeting room as well over the Internet.

The Need for Speed vs. Quality of Decision Making

Integrating AI into office tools speeds up various processes: information is sorted, processed, and presented faster than ever. This can lead businesses to make decisions quicker, riding on the momentum that technology provides.
However, with this increased pace, there’s a danger of reduced reflection time, making decisions which have not been given ample consideration more likely. This will almost certainly lead to poorer outcomes. And the ease with which AI can process and present data might push organisations to act impulsively, with unpredictable, and probably undesirable, results.

Spurious Professionalism

The use of sophisticated AI tools can lend a, perhaps unwarranted, aura of professionalism to presentations and decisions and there’s a real risk that flawed logic or inaccurate data might be glossed over simply because it’s presented in a polished, AI-enhanced manner. We are already biased to see form winning out over content and this is likely to get worse, quickly.
A tool is only as good as its user, and even the most advanced AI can’t make up for a lack of proper review or critical thinking. Relying heavily on AI-driven tools without thorough human oversight can lead to a kind of spurious professionalism—where things look impeccable on the surface but are fundamentally flawed underneath.

In Conclusion…

As AI continues to integrate with office software, businesses will be presented with huge opportunities to improve efficiency and functionality. And competitive pressure will ensure they feel a strong desire to do so. How much real benefit businesses get will depend on thoughtful integration and proper oversight.

Uncategorized

I asked ChatGPT to list 10 good and bad things about LLMs

3rd April 2023 Jim Muttram Leave a comment

Here are 10 ways in which large language models such as myself could provide real benefit to the world:

Facilitating communication: Large language models can be used to bridge language barriers and facilitate communication between people who speak different languages.
Improving healthcare: Language models can help healthcare professionals diagnose and treat patients more accurately and efficiently by analyzing medical data and providing insights.
Enhancing education: Large language models can be used to create interactive and engaging educational materials, such as online courses and textbooks, that adapt to the needs of individual learners.
Boosting scientific research: Language models can assist researchers in analyzing large datasets and identifying patterns and relationships that would be difficult to detect manually.
Improving customer service: Language models can be used to provide 24/7 customer service and support, improving customer satisfaction and reducing wait times.
Enhancing creativity: Language models can assist in generating creative works such as music, art, and literature, allowing for new forms of expression and innovation.
Advancing business: Language models can be used to automate tasks such as data analysis and customer outreach, reducing costs and improving efficiency.
Enabling accessibility: Language models can be used to create more accessible interfaces and products for people with disabilities, including those with visual or hearing impairments.
Facilitating disaster response: Language models can be used to analyze social media data and news reports to provide real-time updates and insights during natural disasters and crises.
Improving personalization: Language models can help businesses and individuals personalize their products and services to meet the unique needs and preferences of their customers.

Here are 10 ways in which large language models could have a detrimental effect:

Bias and discrimination: Language models can perpetuate existing biases and discrimination by reflecting the biases present in the data they are trained on.
Misinformation and fake news: Language models can be used to generate false or misleading information that can spread quickly and easily online.
Privacy and security: Language models can be used to analyze and mine personal data, potentially leading to breaches of privacy and security.
Job displacement: Language models and other forms of AI automation can replace human jobs, leading to economic disruption and social inequality.
Dependence and over-reliance: Dependence on language models and other forms of AI can lead to over-reliance on technology and a lack of critical thinking skills.
Lack of transparency: The algorithms used to train and operate language models can be opaque, leading to a lack of transparency and accountability.
Legal and ethical issues: The use of language models in sensitive areas such as law enforcement and national security can raise legal and ethical concerns.
Loss of human touch: Overreliance on language models and other forms of AI can lead to a loss of human interaction and empathy.
Addiction and social isolation: The use of language models and other forms of technology can contribute to addiction and social isolation.
Environmental impact: The energy consumption and carbon footprint associated with the training and operation of language models and other forms of AI can have a negative impact on the environment.

Interesting……

Internet Economy, Society, Technology

Elon Musk’s strengths and weaknesses

16th December 2022 Jim Muttram Leave a comment

I’ve been thinking about the melt-down which is happening now Elon Musk was reluctantly successful in his purchase of Twitter.

To kill two birds with one stone I thought I would ask ChatGPT what it thought were his strengths and weaknesses:

Elon Musk is a highly successful entrepreneur and CEO, known for his innovative approaches and ambitious projects. However, like all individuals, he also has both strengths and weaknesses.
One of Musk’s greatest strengths is his ability to think big and envision the future. He has a strong vision for what he wants to achieve, and he is not afraid to take risks in pursuit of his goals. This has allowed him to successfully co-found and lead companies such as PayPal, Tesla, and SpaceX, which have all had a significant impact on their respective industries.
Another strength of Musk’s is his ability to adapt and innovate. He is constantly looking for ways to improve and evolve his companies, and he is not afraid to pivot or change course if necessary. This has allowed him to stay ahead of the curve and anticipate market trends, which has been key to the success of his businesses.
However, Musk is not without his weaknesses. One of his most notable weaknesses is his tendency to be overconfident and make bold predictions that may not come to fruition. This has led to some setbacks and criticism when his predictions have not been met. Additionally, Musk’s strong focus on achieving his goals can sometimes lead to him being perceived as lacking empathy or consideration for others.
Overall, Elon Musk is a complex and multifaceted individual with both strengths and weaknesses. While his confidence and vision have helped him achieve tremendous success, it is important for him to continue to strive for balance and consider the impact of his actions on others.

It’s not a bad answer, and exceptional if you think it took 10 seconds for an AI to come up with it.

It’s not a sophisticated answer: it ignored, for instance, his tendency to treat human “laws” (say the desirability of free speech) as if they were the laws of physics, which is his greatest flaw and which has been mercilessly exposed by the shit-show that is Musk-owned Twitter.

But prior to his latest display of his character I would say this is a pretty good assessment. I can only imagine the havoc that will be caused by ChatGPT and its ilk in the very near future!

Climate Change, Electric cars

the joys of vertical integration

25th November 2021 Jim Muttram Leave a comment

One of the many benefits of Tesla vaunted vertical integration is well illustrated by the announcement that off-peak charging will be free at Tesla Superchargers in the California over the Thanksgiving holiday period.

Because Tesla owns all its own charging infrastructure it has a comprehensive view of charging patterns and spikes across its network. This allows it to react to congestion which can be a major problem where there is a large concentration of Tesla EVs.

The company has already tried variable pricing during the day to even out spikes and it has even developed its own mobile Superchargers to boost capacity at high pressure sites. It is able to do this because it makes not only its own chargers but also Megapacks – very large lithium storage batteries which is sells through its power business.

Another reason why Tesla is going to be so hard for automakers to beat.

crypto currency

Sign of the times

26th October 2021 Jim Muttram Leave a comment

How mainstream is the side of a bus? Don’t expect crypto to be there, necessarily. A sign of wider adoption, maybe…

Climate Change, Electric cars

Teslas drive further

29th April 2020 Jim Muttram 1 Comment

No, I’m not talking about range – the perpetual first question in any discussion with non-ev-drivers. I talking about a survey by the RAC, a motoring organisation in the UK, which found that Tesla drivers drive more miles on average than drivers of any other car.

In the first three years of owning a new car, Tesla drivers cover an average of 12,459 miles a year. Meanwhile, Mercedes owners clocked 12,100 miles each year, and Volvo owners averaged 11,578 miles.

This compares to an average of 10,377 miles per year for the average of all cars in their first three years of ownership, according to the Department of Transport.

I can corroborate. I drove an average of 10,000 miles a year in the 10 years I owned a Mercedes E Class. Since I have owned a Tesla Model X I have driven 15,000 in nine months and it would have been more if the Coronavirus hadn’t pretty much put paid to driving.

Electrek, who reported on the RAC survey, concludes: “Electric cars with bigger batteries and faster charging get driven and charged more.”

That is true – but the main thing is they are just so much more fun to drive!

Climate Change, Politics, Science, Society, Uncategorized

Round One: Coronavirus

28th April 2020 Jim Muttram Leave a comment

I took a picture of London from the top of the North Downs in 2017 after I was so struck by the visibility of the pollution hovering over the city. Last week I stopped again at the same spot and took roughly the same photo.

What is so striking is that after only five weeks of lockdown, the dramatic drop in traffic has had such a noticeable visual effect on the air quality.

TomTom, the navigation company, has provided graphs of various cities around the world showing the change in traffic.

What a difference it would make if we could effect a change like this but without the huge downside of a pandemic.

Some cities such as Milan are already planning to reclaim some of their streets inspired by the experience of the traffic-drop. And given that social distancing is likely to be here to stay for quite some time – at least until widespread vaccines are available, others are bound to follow suit.

Wired reports that many cities around the world have already blocked off city streets to provide more open spaces for people to safely navigate.

We could of course go back to normal after the pandemic is over but as The Economist eloquently illustrated coronavirus is merely Round One; the next battle is the big one.

There have been notable examples of self-less co-operation during the coronavirus challenge, but also many examples of narrow-minded, nationalistic responses following the lead of the catastrophically inadequate President of the United States.

We can only hope the sobering example of fighting a pandemic will create real impetus for change which can create a common will to deal with the biggest global challenge of all. Fingers crossed.

Inflection Point